Modelpublic

INCModel3/Z-Image-Turbo-MXFP8-RTN-AutoRound

sourceHugging Faceotherupdated 12d agoView on Hugging Face

0likes16downloads

Model Card

Z-Image-Turbo-MXFP8-RTN-AutoRound

Model Details

This is a MXFP8 (8-bit micro-scaling) quantization of Tongyi-MAI/Z-Image-Turbo, a 6B S3-DiT distilled text-to-image model. Generated by AutoRound with RTN (round-to-nearest, iters=0).

—Base model: Tongyi-MAI/Z-Image-Turbo
—Quantization: MXFP8 (W8A8), group_size=32
—Method: AutoRound RTN
—Model size: ~14 GB (vs 31 GB bf16)

Quantization Details

—Scheme: MXFP8 (datatype=mxfp, bits=8, act_bits=8)
—Group size: 32
—Export format: auto_round (vllm-omni compatible)
—Calibration: coco2014, 8 steps, guidance 0.0
—Ignored layers: adaLN_modulation (kept full precision)

Evaluation

Evaluated with vllm-omni diffusion harness (8 steps, guidance 0.0, 1024×1024, seed 42).

Benchmark	BF16 Baseline	MXFP8 Quantized
DrawBench CLIP	31.73	31.79
DrawBench CLIP-IQA	70.57	70.41
DrawBench ImageReward	1.00	0.97
GenEval	0.757	0.760

MXFP8 quantization is essentially lossless vs the BF16 baseline (GenEval 0.760 vs 0.757, CLIP 31.79 vs 31.73).

Usage

python

from vllm_omni.entrypoints.omni import Omni
from vllm_omni.inputs.data import OmniDiffusionSamplingParams

omni = Omni(model="INCModel3/Z-Image-Turbo-MXFP8-RTN-AutoRound", mode="text-to-image")
params = OmniDiffusionSamplingParams(
    height=1024, width=1024, seed=42,
    guidance_scale=0.0, num_inference_steps=8, num_outputs_per_prompt=1,
)
out = omni.generate("a red bench in a park", sampling_params_list=[params])

License

Please follow the license of the original model Tongyi-MAI/Z-Image-Turbo.

Produced with [autoquant-agent](https://github.com/) — agent-driven quantize + evaluate + self-heal.